GitHub - getmunin/munin: The customer platform for the ag...

MCP-first customer platform made for the agentic era. The agent is the UI.

Munin is an open-source customer platform built for the agentic era (Knowledge Base, Conversations, CRM, CMS, Outreach, Analytics) where the only interface is via AI agents. There is no traditional admin UI for the apps themselves — every action runs through MCP tools, callable from any MCP-compatible client.

See it in action

Lovable builds your frontend. Munin spins up your operations. One prompt, one MCP endpoint — and the agents do the rest.

Watch Lovable build a real website from a single prompt while Munin stands up everything behind it — the CMS the blog reads from, a seeded knowledge base, analytics, and a chat widget that already knows the business. No admin UI, no dashboard to wire up; the agent does the work, over one MCP endpoint. Then a real customer conversation plays out: answered from the knowledge base, handed off to a human when it matters, and picked back up by the agent to close.

Why I built this

Hi — Kjell here. Last year my company was paying HubSpot $500/month for three users who barely touched it. Every workflow worth having lived a tier above the one we were on.

Then our AI agents started doing the prospecting and outreach themselves, and it hit me: what we still used HubSpot for had shrunk to basic CRUD. The agent does the work; HubSpot just stored the record. I wasn't going to keep paying $500/month for CRUD.

So I built the thing I actually wanted. Over roughly a month, with Claude Code as my primary IDE, Munin went from nothing to what's in this repo. I've shipped production software for about twenty years — that experience is what shaped the architecture and let me catch the agent when it got things wrong. Claude Code took the labor of typing it out of the equation; a year ago I don't think this would have been feasible solo.

Munin isn't the AI — it's what your AI uses. You bring the agent, write the prompts, wire up the workflows; Munin's job is to be the cleanest possible tool surface underneath. It's open source and yours to run.

What's in the box

Knowledge Base — markdown documents, hybrid search (BM25 + pgvector embeddings), per-document audience scoping.
Conversations — multi-channel threads (email, voice via Threll.ai or Vapi, SMS via Twilio or MessageBird, chat widget) routed to end-users, assignable, agent-resolvable, with handover state and webhook fan-out.
CRM — contacts, companies, deals, activities, pipelines, segments, plus a merge-proposal queue the clean-contact-data curator runs against.
CMS — agent-authored content collections with field schemas, localized entries, scheduled publishing, an asset library backed by S3-compatible storage, and a public delivery API that ships a _tracking block on every entry for built-in engagement signal.
Outreach — propose-only outbound emails: campaigns + segments + drafts queued for human approval; never auto-sends.
Analytics — polymorphic page-view + search-query ingestion. CMS delivery wires in for free; arbitrary pages drop in a <script src="…/tracker.js" data-key="mn_track_…"> tag. Read tools (analytics_list_top_subjects, analytics_get_subject_engagement, analytics_list_zero_result_searches) feed the "what should we write next" loop into curator skills like cms/review-stale-entries.
Curator — durable background job queue running skills (KB curation, CRM hygiene, contact extraction, stale-content review, outreach drafts) on a schedule with retry + dead-letter.
Playbooks + skills — packaged markdown procedures (skill://module/<verb-object>) the agent reads via MCP resources to follow multi-step flows.
Data portability — symmetric *_export / *_import MCP tools (and /v1/<module>/export|import REST) per module, so an agent can move an org's data between a self-hosted server and the cloud in either direction. See skill://playbooks/data-migration.
Cross-cutting — audit log, webhooks, team invites, OAuth 2.1 dynamic-client registration, BetterAuth-backed sign-in.

Two ways to run

Self-host (this repo): single-tenant, invite-only.

git clone https://github.com/getmunin/munin.git
cd munin
cp .env.example .env  # edit MUNIN_AUTH_SECRET + MUNIN_KEY_PEPPER + MUNIN_ENCRYPTION_KEY
docker compose up

The first user to sign up becomes the org admin; subsequent users need an invitation token or an email whose domain is in MUNIN_ALLOWED_EMAIL_DOMAINS.

Hosted (https://getmunin.com): multi-tenant, one signup per org.

Try it locally

After docker compose up, the backend listens on :3001 and the dashboard on :3000.

Open http://localhost:3000 and register the first user — they become the singleton org admin.
In the dashboard, go to Settings → API keys and mint an admin key (mn_admin_…). Shown once; treat like a password.
Pick one of:

# REST control plane — direct, no OAuth
curl -s http://localhost:3001/v1/kb/spaces \
  -H "Authorization: Bearer mn_admin_..." | jq

# MCP tool browser (recommended for poking at tools/skills)
npx @modelcontextprotocol/inspector
# In its UI: URL = http://localhost:3001/mcp, Auth = Bearer mn_admin_...

# Claude Code (CLI)
claude mcp add munin-local http://localhost:3001/mcp
# Then `claude` → OAuth flow opens in browser → tools available

# Claude Desktop — claude_desktop_config.json
# {
#   "mcpServers": { "munin": { "url": "http://localhost:3001/mcp" } }
# }

# Raw curl over Streamable HTTP — useful for sanity-checking the transport
curl -N -X POST http://localhost:3001/mcp \
  -H "Authorization: Bearer mn_admin_..." \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

The OpenAPI spec for the REST control plane is at packages/backend-core/openapi.json.

Connect your AI agent

Once you've signed up (hosted) or run docker compose up (self-host), point your MCP client at the URL.

Claude Desktop / Code — add to your MCP config:

{
  "mcpServers": {
    "munin": {
      "url": "http://localhost:3001/mcp"
    }
  }
}

(For the hosted version, swap in https://mcp.getmunin.com.) The first call triggers an OAuth consent screen in your browser, then your agent has the full tool surface — Knowledge Base, Conversations, CRM, CMS, Outreach, Analytics.

Documentation

Developer docs live at getmunin.com/docs — guides, the REST API reference, the full MCP tool list, and the skill library.

Two trust contexts, one MCP endpoint

The same /mcp endpoint serves two distinct callers, audience-aware:

Admin agents (Claude Desktop, Cursor, internal automation) — OAuth-authorized by you. Full tool surface, scope-gated per kb:*, conv:*, crm:*, cms:*, outreach:*, analytics:*.
End-user agents (your voice AI, web chatbot, mobile app helper) — short-lived delegated tokens minted server-side from your backend, scoped to one of your end-users. Only self-service tools (read your own contact, send a message in your own conversation).

See packages/backend-core/src/control/delegated-token.controller.ts for the token-mint API. The @getmunin/sdk Node client wraps it.

Architecture sketch

apps/backend — thin NestJS entry composing @getmunin/backend-core modules with single-tenant AuthModule. Exposes the MCP server (Streamable HTTP), OAuth 2.1 + OIDC discovery, and the control-plane REST API on port 3001.
apps/web — Next.js dashboard + landing on port 3000.
apps/chat-widget — embeddable browser widget that consumes a mn_widget_* key and the widget ingest API.
apps/analytics-tracker — embeddable browser tracker, served at /tracker.js, that consumes a public mn_track_* key and writes page-view events.
packages/backend-core — every shared NestJS module (KB, Conversations, CRM, CMS, Outreach, Analytics, Curator, Web, Playbooks, MCP, control plane, audit, tenancy, RLS, mailer, storage, webhooks, rate limit, quotas, OAuth) plus createApp and the createAuthController helper. Cloud composes the same modules with multi-tenant auth.
packages/dashboard-pages — dashboard page components shared between OSS and cloud webs.
packages/ui — design-system primitives (shadcn-style).
packages/{core, db, types, sdk, mcp-toolkit} — non-Nest building blocks: actor identity + tenancy GUCs, Drizzle schema + migrations, shared types, Node client SDK, MCP @McpTool / @SkillRegistry decorators.
packages/{agent-host, agent-runtime} — durable per-org agent runner that picks curator jobs off the queue and executes them against the configured LLM provider.
packages/widget-voice — vendor-agnostic browser voice SDK for the chat widget (Threll.ai WebRTC + Vapi) behind a common VoiceSession interface.
All @getmunin/* packages are published to GitHub Packages.

Stack

TypeScript · Node 24 LTS · Turborepo · pnpm · NestJS · Next.js · Drizzle · Postgres + pgvector · MCP Streamable HTTP · BetterAuth + OAuth 2.1 dynamic client registration

License

MIT. See LICENSE.

Bundled third-party dependencies retain their own licenses — see THIRD_PARTY_LICENSES.md (generated by pnpm licenses:generate, verified in CI).

推荐订阅源

Hacker News: Show HN